JP6444977B2 - Device for controlling self-driving or partially self-driving land vehicles - Google Patents

Description

  The present invention relates to a control device, and more particularly, to an apparatus for controlling a land vehicle that is self-driving or partially self-driving.

  The realization of self-driving vehicles or partially self-driving vehicles is complex, depending on the driver's degree of involvement. Requirements include lane keeping and warning system support, adaptive cruise control, backup warnings, and parking assistance. Advanced driver assistance systems (ADAS), light detection and ranging (LIDAR), and dedicated short-range communications (DSRC) in real-time inter-vehicle communications are just some of the known technologies under development. These techniques are designed to allow the car to calculate the 3D geometry of any situation in real time from the image and respond to these surroundings. For example, ADAS uses a combination of altitude sensors, such as stereo cameras and long and short range radars, combined with actuators, control units, and integrated software.

  So far, the fusion of available sensors and artificial intelligence has not been able to "see" and understand the situation around the vehicle as accurately as humans can. Artificial intelligence could not provide that level of reasoning thinking and could not communicate with the environment in real time. It has been realized that concentrating these sensors and connection technologies increases the input available for decision making and reduces the need for sophisticated artificial intelligence.

  Model predictive control (MPC) is an optimization-based control strategy and is considered very attractive in vehicle autonomous driving scenarios. For example, for MPC solutions, APACC uses sensors, data including driver behavior and traffic data, and model predictive controllers to improve fuel economy.

  Safe self-driving vehicles have already been made. Google, as well as traditional car manufacturers and suppliers, has developed a self-driving function using sensor-based solutions and has many new applications in progress. At the same time, multiple organizations, including automotive and high-tech companies and USDOT, are focusing on the potential of using connected vehicle communication technology for collision avoidance and traffic management.

  Integrating sensor-based technology and connected vehicle communications supported by advanced control schemes is required to enable true autonomous vehicles. However, in the advanced control system of the past several years, linear model predictive control (MPC) has only been used in vehicle autonomous driving. The basic idea behind all MPC techniques is to rely on plant model predictions to calculate the optimal future control sequence by minimizing the target function. The MPC model includes controlled variables, manipulated variables, and disturbance (perturbation) variables. At each sample time point, the optimization performed based on the new measurement data and the initial control input of the sequence is applied. The rest of the sequence is discarded and the process is repeated at the next sampling time in a “backward horizon” manner.

  Although MPC has its origins in the chemical process industry, it is increasingly interested in its application to vehicle autonomous driving, including support for lane keeping and warning systems, adaptive cruise control, backup warnings, and parking assistance. In essence, MFC applications are built on the philosophy of fuel and time optimal speed planning by adopting optimization and providing a natural framework for increasing autonomy and reconfiguration, while providing finite control authority ( It takes into account physical and operational constraints such as finite control authority, passive safety, and collision avoidance. In addition, other improvements are possible, such as an itinerary that is otherwise unknown.

  Generalized predictive control (GPC) and its derivatives have received special attention. In particular, GPC can be applied in a simple manner to an unstable or time-delayed MIMO system, and it is useful for many different types of tasks because of the low computational requirements for stationary models. Unpredictable systems, human factors, faults, and the environment are all contributing factors to nonlinear plant dynamics. The disadvantage of MPC is that this method is limited to linear models. If non-linear dynamics are present in the plant, the linear model may not be able to make sufficient predictions for the MPC technique to function properly.

  The two most common approaches to closed loop identification are the direct approach and the indirect approach. The direct approach ignores the presence of feedback and identifies the plant by direct plant input and output data. This has the advantage that no knowledge of the type of control feedback or controller linearity is required. The indirect approach identifies closed loops and obtains open loop models by deconvolution where possible. An open loop model can be obtained only if the controller is known and both the closed loop plant model and the controller are linear.

  Control of multi-variable multiple-input multiple-output (MIMO) systems is a common problem in actual vehicle-to-vehicle communication and communication from vehicle to infrastructure. Connected vehicle systems use wireless technology to communicate in real time. Most existing approaches only deal with driver behavior and environmental uncertainties, assuming that an accurate model of the actuator is available. It is rare that this assumption is met. This is because actuator parameters also have uncertainties due to input disturbances such as mechanical and electrical components. Although adaptive control with actuator uncertainty is not considered, this uncertainty results in significant degradation in controller performance.

  The object of the present invention is to reduce the problems mentioned above.

  In one non-limiting embodiment of the present invention, an apparatus for controlling a self-driving or partially self-driving land vehicle is provided. The apparatus includes a coarse adjustment assembly and a fine adjustment assembly.

The coarse adjustment assembly
a. A sensor interface for measuring mechanical parameters including speed and braking;
b. A fuzzy description that models vehicle guidance, navigation, and control,
(I) the driver's behavior and driving dynamics;
(Ii) environmental uncertainty including weather, road conditions, and traffic volume;
(Iii) input disturbances including mechanical and electrical components;
Including fuzzy description,
c. An adaptive fuzzy logic controller for a nonlinear MIMO system with subsystems that includes fuzzification, inference, and output processing including type reduction and defuzzification, and provides stability to the resulting closed-loop system. And
(I) an inference engine that identifies the relationship using the rule base and output as a “fuzzy set” for the type reducer;
(Ii) an adaptive fuzzy logic controller including an output control demand including from a torque actuator to a fuzzifier that “fuzzifies” the signal;
Is included, and
The fine adjustment assembly
a. Input from the coarse adjustment assembly;
b. A predictive horizon that determines how many future samples the objective function considers for minimization and length of the calculated control sequence;
c. A linearized MIMO regression model extracted from an adaptive fuzzy logic controller for each time step supplying a “fine” adjustment parameter;
d. A non-linear dynamic linearization regression controller comprising:
(I) a crisp output signal that is input to an APACC integration that calculates an optimal future vehicle guidance, navigation, and control sequence;
(Ii) a non-linear dynamic linearization regression controller that provides reduced set output and APACC integration for input to the APACC linear logic system.

  The apparatus may include a synchronization assembly that includes a cascade Diophantine frequency synthesis (DFS) means that optimizes the input signal to the output signal and predicts the vehicle's future stabilized output parameters.

The use of neural networks for system identification is a relatively new approach, as is fuzzy logic.
The device of the present invention can be considered to provide artificial prediction (AP) using adaptive (model) cognitive control (APACC). APACC improves model predictive control (MPC). Using a very complex algorithm, APACC allows for a very high level of reasoning thinking and real-time communication. With the integration of available sensors, APACC provides the ability to “perceive” the surroundings of the vehicle and give a true “perception” of the external environment. This is somewhat helpful in simulating how humans can use a combination of memory and sensory inputs to interpret these as events occur and to predict (recognize) possible scenarios. By surpassing what has ever been possible.

Embodiments of the present invention will now be described by way of example only and with reference to the accompanying drawings.
FIG. 1 shows APACC in vehicle autonomous driving. FIG. 2 shows a computer hardware core system for artificial prediction using APACC in vehicle autonomous driving. FIG. 3 shows an Intel® Xeon® processor that performs high-performance calculations for APACC in vehicle autonomous operation.

  Referring to the drawings, non-linear vehicle autonomous driving dynamics can be characterized as “fuzzy” with a high degree of non-linearity. APACC provides the “cognitive” output of the fuzzy logic circuit (fuzzyifier in FIG. 2) at each sampling to the instantaneous linearity of the nonlinear model. This is similar to GPC in most aspects, except that the instantaneous linearity of the fuzzy logic output produces an adaptive linear regression model.

  The main advantage of fuzzy logic is that it allows the designer to describe the desired system behavior in a simple “if-then” relationship. In many applications, this results in a simpler solution with less design time. In addition, the designer can use all available design know-how to directly optimize system performance. This is certainly the beauty of fuzzy logic, but it also has serious drawbacks. In many applications, knowledge describing the desired system behavior is contained in the data set. Here, the designer has to manually derive the “if-then” rule from the data set, but a large data set requires a great deal of labor. When the data set contains knowledge about the system being designed, the neural net promises a solution because it can train itself from this data set.

  Neural nets have the advantage of learning from a data set, but they have inherent disadvantages. For example, the cause for a particular behavior cannot be interpreted, and the neural net cannot be changed manually to change to a desired behavior. In addition, selection of appropriate net models and setting of learning algorithm parameters are difficult and require a lot of experience. On the other hand, the fuzzy logic solution is easy to verify and optimize. The present invention utilizes a fuzzy logic controller that automates rule derivation and eliminates the need to manually perform this function to instantaneously predict plant dynamics.

  Fuzzy control methods have emerged in recent years as a promising way to tackle nonlinear control problems. In particular, fuzzy control has influenced the world of control because of the simple approach it provides to use heuristic control knowledge for nonlinear control problems. In very complex situations, if the plant parameters are perturbed, or if the dynamics of the system are very complex, use an adaptive method online to collect data and automatically adjust the control parameters Must. However, these adaptation techniques have not been given stability conditions so far. APACC introduces two components into its adaptive fuzzy control scheme. One is a coarse adjustment fuzzy logic system. The other is instantaneous linearization of the fuzzy logic circuit output that generates an adaptive linear model. This functions as a kind of robust compensator, such as supervisory control for fine adjustment and sliding mode control.

  In recent years, various stable adaptive fuzzy control schemes have been developed for multiple input multiple output (MIMO) nonlinear systems. However, these adaptive control techniques are limited only to MIMO non-linear systems where states are assumed to be available for measurement. In many practical situations, state variables are often not available in nonlinear systems. That is, in such complicated applications, output feedback or APAGC adaptive fuzzy control is required. The fuzzy control system controls the MIMO system and maintains system stability. Coarse and fine adjustments improve system performance by reducing the effects of external perturbations and ensuring closed loop stability.

  APAGC coarse and fine adjustments apply to land vehicles, for example, full or partial autonomous control of vehicles. These include dynamics such as those provided by 3D scene structure, speed, braking, and tire sensor and actuator input uncertainties to instantly recognize possible 3D scene structure / mechanical / actuator parameters Applied to variables.

  Non-adaptive passive methods applied to vehicle autonomous driving are uncertainties in “linear” (eg, mechanical relationships) and “non-linear” (eg, driver behavior, environmental, input disturbances in mechanical and electrical components) And an accurate model of the actuator is available.

  As shown in FIG. 1, there are three levels of schematic diagrams. Vehicle-mounted cameras, sensors, ADAS, LIDAR, and DSRC in the inter-vehicle and vehicle / infrastructure technology shown in FIG. 1 generate specific real-time vehicle data. This vehicle data is combined with actuators, control units, and integrated software to allow the vehicle to monitor and respond to its surroundings. These sensor-based, driver-assisted solutions using 3D or stereo cameras can calculate 3D geometries for any situation in front of the vehicle in real time from images viewed by the vehicle. Vehicle data is combined in real time with driver behavior, environment (eg, weather, road conditions, traffic) maps, and images stored in a high performance computing (HPC) data center. This provides a “perception” of the external environment, and the fusion of available sensors and APACC provides the ability to “recognize” the surroundings of the vehicle. This uses a combination of human-stored memory and sensor input to mimic how to interpret these as events occur and predict possible scenarios. Using a very complex algorithm, APACC provides a level of inference thinking and real-time communication.

  FIG. 2 shows how APACC is used for vehicle autonomous driving. Designs for controlling vehicle dynamics include sensors, actuators, and (1) sensor measurements of current vehicle dynamics, and (2) algorithms for commanding actuators based on desired dynamics specifications Need.

  Fully and partially autonomous vehicles have outputs corresponding to automatic dynamic motion as a result of application of APACC. Components essential for artificial prediction are embedded as a subsystem in the data center implementation example 1 of the high-performance computer hardware core system shown in FIG. A detailed implementation example of the core system is shown in FIG. Input includes perturbations to the vehicle. Input includes a large number of variables, including speed, braking, and dynamic variables such as those supplied by tire sensor and actuator input uncertainties. Input / output is managed by a multivariate multiple input multiple output (MIMO) subsystem 2 as shown in FIG. The sensor delta is measured by the sensor through a sequential control process. The sequential control process includes feedback measurements (feedback elements) and a comparator that compares the differences between the input and output signals. The MIMO subsystem 2 is driven to different operating states using set points that are added to the input signal at the comparator. The actual excitation signal amplitude should be as large as possible to ensure maximum excitation around each set point. APACC instantly compensates for errors between actual power and perceived (possible and likely) power (eg throttle control and maneuvering) to provide stability to the vehicle dynamics . This error is shown in the MIMO subsystem 2 to be supplied to a controller that provides a control output to the actuator. Sensor actuators add the mechanical action needed to change the direction or speed of the vehicle.

  Coarse adjustment is performed by the adaptive fuzzy logic controller 3 shown in FIG. It was developed to model car guidance, navigation, and control using fuzzy descriptions for nonlinear MIMO systems with external perturbations. An adaptive fuzzy logic controller is a type-2 fuzzy logic system with fuzzification, inference, and output processing operations. The sensor / actuator outputs a control request to a fuzzer that “fuzzifies” the signal. Some of the inputs are closely correlated variables (described above) that have high mutual information. However, some other related variable pairs have low correlation but high mutual information. Coarse adjustment drastically speeds up the cognitive process and allows for virtually instantaneous dynamic correction. The inference engine uses the rule base to identify these relationships and outputs them as a “fuzzy set” to the type reducer. The adaptive fuzzy logic controller performs “output processing” and includes both type reduction and defuzzification.

  Type reduction (reduced set) captures more information about rule uncertainty than defuzzified values (crisp number), but is computationally intensive. The advantage is the ability to recognize unpredicted perturbations and data uncertainties. The adaptive fuzzy controller can perform perfect control and ensure that the resulting closed loop system and overall stability of the tracking performance can be achieved.

  The output processing of the adaptive fuzzy logic controller is fed to the nonlinear dynamic linearization regression controller 4 shown in FIG. The controller 4 performs fine adjustment, and the crisp output signal is input to APACC synthesis. Reduced set output and APACC integration input to the APACC linear logic system. A linearized MIMO regression model is extracted from the fuzzy logic controller at each time step and used to provide “fine” adjustment parameters. The non-linear dynamic linearization regression controller calculates optimal future guidance, navigation, and control sequences according to the objective function. Due to the minimization of the calculated control sequence and its length, two “predict horizons” determine how many future samples the objective function considers. As is common in most MPC methods, a reverse horizon strategy is used, ie, only the first calculated control signal is actually applied to the vehicle system to achieve loop closure.

  Using cascaded diophantine frequency synthesis (DFS), where synchronization optimization of the input signal to the output signal is achieved using two or more phase-locked loops (PLLs) Achieved. The DFS is the DFS 5 in FIG. DFS5 is used to predict the future output of continuous time, infinite dimension, time variable, and nonlinear systems. Its primary function is to parameterize vehicle stabilization factors and lock them into a continuous feedback loop.

  APACC for land vehicle guidance, navigation, and control (GNC) requires virtually instantaneous analysis of vast amounts of data. To accomplish this, a conveyor high performance computing (HPC) architecture from Intel® was selected for APACC vehicle autonomous driving. The Combay computer approach allows very fast access to random access memory and is very useful for complex functions used in APACC.

  This architecture is based on the Intel® Xeon® processor shown in FIG. This architecture features a highly parallel memory subsystem to further improve performance. Programmable “on the fly” FPGAs are one way to achieve hardware-based application specific performance. A particular APACC-GNC algorithm is, for example, optimized and converted into code, which is loaded into the FPGA at runtime.

  Vehicle-mounted cameras, sensors, ADAS, LIDAR, and DSRC in vehicle-to-vehicle and vehicle-infrastructure technology present inputs to APACC for full or partial vehicle autonomous driving.

APACC assembly is
a. A sensor for measuring vehicle dynamics;
(I) an input surface that is locally exposed to vehicle perturbations;
(Ii) input parameters include those supplied by mechanical parameters and tire sensor and actuator input uncertainties;
b. An actuator for applying operations including steering, throttle and braking to change the course of the vehicle;
c. (1) current course sensor measurements, and (2) an algorithm to command the actuator based on the desired course specifications;
including.

  APACC mimics how humans use a combination of stored memory and sensory input to interpret events as events occur and to predict (recognize) possible scenarios.

  The multivariable multiple input multiple output (MIMO) subsystem 2 includes a sensor assembly. The sensor assembly essentially includes a sequential control process that measures feedback and a comparator that compares the difference between the input and output signals. The MIMO subsystem 2 operates as follows.

a. The set point added to the input signal in the comparator is used to drive the MIMO subsystem 2 to different operating states.
b. APACC instantaneously compensates for errors between the actual output and the perceived (possible and likely) output. The output includes throttle control and steering that provides stability to the vehicle dynamics.

c. The excitation signal amplitude is maximized to ensure maximum excitation around each set point.
d. It provides errors to the controller that provides the control output to the sensor / actuator and provides the mechanical action necessary to change the direction or speed of the vehicle.

  A high performance computing (HPC) architecture allows very fast access to random access memory for virtually instantaneous analysis of vast amounts of data for car guidance, navigation, and control.

  It will be appreciated that the embodiments of the present invention described above with reference to the accompanying drawings have been given by way of example and with reference to the accompanying drawings. The individual components shown in the drawings are not limited to use in the drawings, and can be used in other drawings and all forms of the invention.

Claims (2)

Self-driving or partially self-driving land vehicle control device comprising a coarse adjustment assembly and a fine adjustment assembly;
The coarse adjustment assembly comprises:
a. A sensor interface for measuring mechanical parameters including speed and braking;
b. A fuzzy description modeling the guidance, navigation and control of the vehicle,
(I) the driver's behavior and driving dynamics;
(Ii) environmental uncertainty including weather, road conditions, and traffic volume;
(Iii) input disturbances including mechanical and electrical components;
Including fuzzy description,
c. An adaptive fuzzy logic controller for a nonlinear MIMO system with subsystems that includes fuzzification, inference, and output processing including type reduction and defuzzification, and provides stability to the resulting closed-loop system. And
(I) an inference engine that identifies the relationship using the rule base and output as a “fuzzy set” for the type reducer;
(Ii) an output control request including from a torque actuator to a fuzzifier that “fuzzifies” the signal;
An adaptive fuzzy logic controller, including
Is included, and
The fine adjustment assembly comprises:
a. Input from the coarse adjustment assembly;
b. A predictive horizon that determines how many future samples the objective function considers for minimization and length of the calculated control sequence;
c. A linearized MIMO regression model extracted from the adaptive fuzzy logic controller for each time step supplying the “fine” adjustment parameter;
d. A non-linear dynamic linearization regression controller that calculates optimal future vehicle guidance, navigation, and control sequences ,
(I) crisp output signal and APACC integration provides input,
(Ii) reduction set output and APACC integration and APACC linear logic system provides input,
A nonlinear dynamic linearization regression controller comprising:
A device that is supposed to contain.   2. The apparatus of claim 1, comprising cascade Diophantine frequency synthesis (DFS) means for optimizing the input signal with respect to the output signal and predicting future stabilized output parameters of the vehicle. A device comprising a synchronization assembly.

Images